I wrote: > it might be worthwhile to "zoom in" with more measurements, but haven't done that yet.
I've attached the script and image for 1 million / random / varying the mod by quarter-log intervals. Unfortunately I didn't get as good results as yesterday. Immediately going from mod 1 to mod 2, sort pushdown regresses sharply and stays regressed up until 10000. The tiebreaker patch helps but never removes the regression. I suspect that I fat-fingered work_mem yesterday, so next I'll pick a badly-performing mod like 32, then range over work_mem and see if that explains anything, especially whether L3 effects are in fact more important in this workload. -- John Naylor EDB: http://www.enterprisedb.com
bench_windowsort-jcn-random-finegrained.sh
Description: application/shellscript